{"slug": "making-process-cheap", "title": "Making Process Cheap", "summary": "An Anthropic engineer stated that production code written by Claude should meet a higher bar than human-written code, citing lint rules, tests, Claude-driven end-to-end tests, Claude-powered fuzzers running daily, automated code and security reviews, and automated refactoring as guardrails, a statement Simon Willison highlighted. The commentary argues teams can keep process and speed with LLM coding agents by applying them to micro-level topic branches and PRs, nightly CI-style jobs, and less-frequent analyses, with local models suited to smaller scopes and overnight checks.", "body_md": "Simon Willison [pointed out a statement by an Anthropic engineer](https://simonwillison.net/2026/Sep/11/boris-cherny/):\n\nProduction code written by Claude should have a higher bar than if it was written by a human. At Anthropic, we have many guardrails in place to make sure this is happening: lots of lint rules, lots of tests, Claude-driven end to end tests, Claude-powered fuzzers running daily, automated code reviews and security reviews, automated code refactoring, and so on. Without these, you can end up with a mess that is hard to maintain down the line.\n\nI am reasonably certain that Anthropic does this sort of work using agent swarms. I am even more certain that Anthropic cares very little about internal token spend.\n\nThat said, the point is valid: **we should be using LLMs to improve what we are delivering above the traditional baseline**. \"Same old, same old\" but a lot faster IMHO is not the right answer. I prefer a little bit faster but with substantially better results. We can use coding agents to accelerate the development process, tackling the things that we might be inclined to skip in order to save engineering time.\n\nSome improvements can be made at a \"micro\" level, focusing on an individual topic branch or PR, where you have the coding agent:\n\n`key`)\nOther concerns might get surfaced via nightly CI-style jobs, applied against specific modules for large projects if not the entire project, such as:\n\nStill more analyses could be done on a less-frequent basis, such as:\n\nAnd, from time to time, coding agents can help improve the overall development process for natural and artificial developers, such as crafting new compile-time checks (lint, Detekt, etc.), distilling knowledge into new skills, and so on.\n\nMany of these will not necessarily require frontier models, though larger projects will be more likely to need the sheer speed that frontier models offer. If you want to emphasize local models, do more analyses on smaller scopes: the single topic branch/PR, the single module, etc. Rig up the local model to do overnight checks (Monday for security, Tuesday for performance, Wednesday for accessibility, etc.) and dump a report for you to read in the morning.\n\nSome teams think that process needs to be thrown out the window to achieve speed. In a world of LLM-powered coding agents, we can have both process and speed, so long as we actually pay attention to the process part.", "url": "https://wpnews.pro/news/making-process-cheap", "canonical_source": "https://pac.commonsware.com/archive/making-process-cheap/", "published_at": "2026-09-12 20:24:53+00:00", "updated_at": "2026-09-12 20:55:43.593739+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "developer-tools", "ai-products"], "entities": ["Anthropic", "Claude", "Simon Willison", "Boris Cherny", "Detekt"], "alternates": {"html": "https://wpnews.pro/news/making-process-cheap", "markdown": "https://wpnews.pro/news/making-process-cheap.md", "text": "https://wpnews.pro/news/making-process-cheap.txt", "jsonld": "https://wpnews.pro/news/making-process-cheap.jsonld"}}